💬 NLP

Dual Perspectives in Emotion Attribution: A Generator-Interpreter Framework for Cross-Cultural Analysis of Emotion in LLMs

This paper proposes a Generator-Interpreter framework to address the overlooked cultural background of emotion expression in large language models, revealing through a cross-cultural evaluation of 15 countries that both expression and interpretation perspectives significantly impact emotion attribution performance, with the generator's origin playing a dominant role.

Original authors: Aizirek Turdubaeva, Uichin Lee

Published 2026-04-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Aizirek Turdubaeva, Uichin Lee

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a huge international potluck dinner. Everyone brings a dish that represents their home, but there's a twist: no one speaks the same language, and everyone has a different idea of what "spicy," "sweet," or "sour" actually means.

This paper is about what happens when Artificial Intelligence (AI) tries to taste these dishes and guess what the cook was feeling when they made them.

Here is the breakdown of the research in simple terms:

1. The Big Problem: The "One-Way Street" Mistake

For a long time, researchers thought AI emotion detection was a one-way street. They asked: "If I show this AI a sad story, will it correctly say 'Sadness'?"

But the authors realized this is like judging a movie only by the director's script, ignoring the actor's performance. They argued that emotion has two sides:

The Generator (The Cook): The person who wrote the story. They grew up in a specific culture with specific rules about how to show feelings. (e.g., In some cultures, you hide anger; in others, you shout it).
The Interpreter (The Taster): The AI (or a person) trying to guess the feeling. They also have a cultural background that shapes how they read the story.

The Analogy: Imagine a Swiss person writes a story about feeling "ashamed." In their culture, this might be a quiet, internal feeling. If an American AI reads it, the AI might think, "Oh, this person is just feeling guilty about a mistake," because that's how Americans often interpret shame. The AI misses the cultural nuance.

2. The Experiment: The "Passport Swap"

The researchers set up a massive experiment using data from 15 different countries (like China, Brazil, India, USA, etc.) and 6 different AI models (like GPT-4, Claude, DeepSeek).

They played a game of "Passport Swap":

They took a story written by someone from China (The Generator).
They asked the AI to pretend to be a person from Zambia (The Interpreter) and guess the emotion.
Then they asked the AI to pretend to be a person from Austria and guess the same story.
They did this for every possible combination (15 countries × 15 countries × 6 AIs).

Total moves: Nearly half a million guesses!

3. What They Found: The "Cultural Blind Spot"

The results were eye-opening. Here are the main takeaways:

The "Origin" Matters More Than the "Reader":
The biggest factor in whether the AI got it right wasn't which AI was reading the story, but where the story came from.
- Analogy: It's like trying to read a book written in a dialect you've never heard. Even if you are a super-smart reader (a top-tier AI), if the book was written in a cultural style you don't understand, you will still get confused.
- Stories from China and Zambia were consistently the hardest for all AIs to understand, no matter which AI persona was doing the reading.
The "Sadness" Trap:
The AIs had a habit of turning almost everything into "Sadness."
- If a story was actually about Anger, Disgust, or Shame, the AI often just said, "This is Sadness."
- Analogy: It's like a colorblind painter who sees everything as shades of gray. They can't distinguish the bright red of anger from the deep blue of sadness, so they just paint it all blue.
The "Shame vs. Guilt" Confusion:
In many Western cultures, Guilt is about "I did something bad." In many Eastern/Collectivist cultures, Shame is about "I made my family/community look bad."
- The AIs struggled to tell these apart. They often swapped them, especially when reading stories from collectivist cultures.
Does "Same-Country" Help?
You might think, "If the AI pretends to be Chinese, it will understand Chinese stories better."
- The Result: Sometimes, yes. But not always. Even when the AI pretended to be from the same country as the writer, it still struggled with certain complex emotions. The "cultural gap" is deeper than just pretending to have a passport.

4. Why This Matters for the Future

We are building AI assistants for mental health, customer service, and education. These AIs need to understand how people feel to be helpful.

The Risk: If an AI thinks a person is just "sad" when they are actually "angry" because of a cultural misunderstanding, it might give the wrong advice. It might tell an angry person to "calm down" instead of helping them solve the injustice they are facing.
The Solution: The authors say we need to stop treating AI as a universal translator. We need to build systems that ask: "Who wrote this? Where are they from? And who is reading it?"

The Bottom Line

Emotions aren't universal math problems (1 + 1 = 2). They are more like dialects.

If you want an AI to truly understand human feelings, you can't just teach it the dictionary. You have to teach it the culture, the history, and the context of the person speaking. Otherwise, the AI is just guessing, and in the world of emotions, a wrong guess can feel like a betrayal.

1. Problem Statement

Current Large Language Models (LLMs) are increasingly deployed in cross-cultural systems for social-emotional support, mental health, and education. However, existing research on emotion attribution in LLMs suffers from a critical limitation: it predominantly focuses on the interpreter's perspective (how a model classifies emotions in text) while neglecting the generator's cultural background (the origin of the emotional narrative).

This "one-way" assumption of universality ignores the fact that emotions are culturally mediated. The same emotional event may be expressed differently depending on cultural display rules (generation) and interpreted differently based on cultural attribution theories (interpretation). Consequently, current models often exhibit Western-centric biases, misinterpreting emotions from non-Western cultures, leading to cultural misalignment and reduced effectiveness in global applications.

2. Methodology

Framework: Generator-Interpreter

The authors propose a novel Generator-Interpreter Framework that explicitly models the dual nature of emotion attribution:

Generator: The cultural origin of the emotional event (e.g., a narrative written by a Chinese participant).
Interpreter: The cultural persona adopted by the LLM to classify the emotion (e.g., "Act as an Indian annotator").

Dataset

Source: The ISEAR (International Survey on Emotion Antecedents and Reactions) dataset.
Scale: 5,250 emotional events across 15 countries (Australia, Austria, Brazil, Bulgaria, China, Finland, Honduras, India, Malawi, Netherlands, New Zealand, Spain, Sweden, USA, Zambia).
Emotions: Seven categories: Joy, Fear, Anger, Sadness, Disgust, Shame, and Guilt.
Preprocessing: The dataset was cleaned to remove invalid responses and balanced to 350 events per country. Although the text is in English, the cultural origin of the narrative is preserved to isolate cultural framing from linguistic nuances.

Models Evaluated

Six state-of-the-art LLMs were tested, representing a mix of Western and Eastern development origins:

Western: GPT-4o, Gemini, Claude.
Eastern: Qwen2, DeepSeek.
Open Source: Mistral.

Experimental Design

Task: Zero-shot emotion attribution. The model is given an event and a specific nationality persona, then asked to identify the primary emotion.
Scope: A full factorial design where every event (15 countries) is evaluated by every persona (15 countries) across all 6 models.
Total Attributions: 472,500 ( $5,250 \text{ events} \times 15 \text{ personas} \times 6 \text{ models}$ ).
Evaluation: Accuracy, Precision, Recall, and F1-score were calculated using bootstrap resampling (100 iterations).

3. Key Contributions

Theoretical Framework: Introduction of the Generator $\times$ Interpreter evaluation framework, shifting the paradigm from unidirectional emotion recognition to a bidirectional analysis of cultural origin and target.
Systematic Empirical Study: The first large-scale evaluation of cross-cultural emotion attribution across 15 nationalities and 6 diverse LLMs, revealing specific biases and alignment effects.
Identification of Asymmetries: Discovery that the generator's country of origin has a stronger impact on attribution difficulty than the interpreter's persona, challenging the assumption that model performance is solely dependent on the interpreter's cultural alignment.

4. Key Results

Performance Variations by Emotion

Robust Emotions: Joy and Fear were consistently recognized with high accuracy (F1 > 0.93 for Joy) across all models and cultures.
Challenging Emotions: Shame, Guilt, and Disgust were the most difficult to detect (F1 often < 0.65).
- Misclassification Patterns: Shame and guilt are frequently confused with each other, particularly in collectivist contexts. Sadness acts as a "default" category; models often collapse Anger, Disgust, and Shame into Sadness.
- Cultural Nuance: Models tend to assign "Shame" more frequently to Chinese personas (reflecting collectivist social cohesion) and "Guilt" to individualist societies, reproducing existing cultural stereotypes.

Generator-Interpreter Alignment Effects

Alignment is Not Universal: While same-country alignment (e.g., Chinese generator + Chinese interpreter) sometimes improved accuracy, it did not guarantee superior performance across all models.
Generator Dominance: The difficulty of interpreting an emotion is heavily determined by the generator's origin.
- Events generated by China and Zambia were systematically harder for all interpreters (including their own) to classify correctly. For example, GPT-4o dropped to ~0.62 accuracy on Chinese-generated events compared to ~0.75 on Western events.
- Conversely, events from Austria, Brazil, and Malawi were interpreted more reliably across the board.
Interpreter Capability: Some personas (USA, Austria, Spain, Bulgaria) consistently achieved higher accuracy as interpreters, suggesting certain cultural frameworks are more "universal" or better represented in training data.

Qualitative Findings

Self-Conscious Emotions: In Chinese-origin events, "Shame" (social evaluation) was often misinterpreted by non-Chinese personas as "Guilt" (individual moral failure).
Anger Attenuation: Anger in victimization scenarios (e.g., sexual assault, public insult) was frequently attenuated into inward-oriented emotions like Sadness or Shame by non-native interpreters, failing to recognize the moral outrage.
Moral Disgust: Cultural or moral disgust (e.g., political or social norms) was often misclassified as Joy or neutral, indicating models struggle with culturally embedded moral aversion.

5. Significance and Implications

Beyond "Interpreter-Only" Benchmarks: The study proves that evaluating LLMs solely on their ability to interpret text is insufficient. The cultural origin of the input data is a critical variable that dictates performance.
Bias Mitigation: Current LLMs are biased toward Western, individualist frameworks. This leads to systematic errors when processing emotions from collectivist cultures, potentially causing harm in mental health or educational applications.
Design Recommendations:
1. Contextual Awareness: Systems should infer or explicitly request the user's cultural background (location, language, style) to condition the interpreter persona.
2. Data Diversity: Training data must be expanded to include diverse emotional expressions from low-resource and non-Western cultures to reduce the "generator difficulty" gap.
3. Evaluation Standards: Future benchmarks must adopt the Generator $\times$ Interpreter matrix to detect cross-cultural misalignment rather than relying on aggregate accuracy.

In conclusion, the paper argues that for LLMs to be truly empathetic and effective in a global context, they must be designed with culture-aware generation and interpretation, acknowledging that emotion is not a universal signal but a culturally constructed phenomenon.